Search Results

Jakob Hohwy seeks to recover an approach to consciousness from a general theory of brain function, namely the prediction error minimization theory. The way this theory applies to mental and developmental disorder demonstrates its relevance to consciousness. The resulting view is discussed in relation to a contemporary theory of consciousness, namely, the idea that conscious perception depends on Bayesian metacognition which is also supported by considerations of psychopathology. This Bayesian theory is first disconnected from the higher-order thought theory, and then, via a prediction error conception of action, connected instead to the global workspace theory. Considerations of mental and developmental disorder therefore show that a very general theory of brain function is relevant to explaining the structure of conscious perception. Furthermore, Hohwy argues that this theory can unify two contemporary approaches to consciousness in a move that seeks to elucidate the fundamental mechanism for the selection of representational content into consciousness.

In numerous brain areas, neuronal activity varies according to reward predictability. In many of these areas this activity is thought to represent errors in reward prediction, as has been described for dopamine neurons; however, it might alternatively be related to the animal's enhanced behavioural state, which may include surprise or changes in arousal or attention. Unfortunately, few studies have examined firing in these areas in the context of the same behavioural task, making it difficult to dissociate different types of encoding. This chapter compares neural correlates in these areas to that of dopamine neurons in the same behavioural task. It shows that while activity in dopamine neurons appears to signal prediction errors, similar activity in the orbitofrontal cortex, basolateral amygdala, and ventral striatum does not. Instead, increased firing in basolateral amygdala to unexpected outcomes likely reflects attention, whereas activity in the orbitofrontal cortex and ventral striatum is unaffected by prior expectations and may provide information on outcome expectancy. These results have important implications for how these areas interact to facilitate learning and guide behaviour.

This part presents four chapters on the concept of learning. The first chapter argues that learning is a pre-theoretical, possibly even a pre-scientific, concept. The second discusses three concepts that are central to more than one theory: associative strength, associability, and prediction error. The third focuses on challenges in linking the primary, behavioral indication of learning with cellular and systems neurobiological approaches to learning mechanisms. The fourth chapter presents a synthesis of the chatpers in this part.

The central mechanism for hierarchical perceptual inference is prediction on the basis of internal, generative models, revision of model parameters, and minimization of prediction error. This is the way in which the brain engages in perceptual inference, as described in the previous chapter. This chapter describes this idea in detail. It uses a statistical analogy of model fitting to explain the notion of prediction error and then gradually builds up more and more complex versions of the theory, ending with broad ideas from information theory and statistical physics concerning mutual information, free energy and surprisal. The overall picture is of a self-supervised system that is closely supervised by the sensory signal it receives from the world, but which is hidden behind the veil of sensory input. This is a profound reversal of the way we normally think about the top-down and bottom-up signals in the brain. The system is able to recognize the causes of its sensory input in a mechanistic manner, by implicitly inverting its generative model.

This chapter considers some very simple learning problems to make accurate predictions. It reviews the least mean squared (LMS) algorithm. It shows that internal model is simply a link between motor commands and their sensory consequences. The driving force in learning an internal model is the sensory prediction error. This chapter also reveals that when motor commands are generated, perturbations like force fields or visuomotor rotations produce discrepancies between the predicted and observed sensory consequences. It illustrates that in some forms of biological learning, as in backward blocking, animals seem to learn in a way that resembles the Bayesian method and not LMS.

Psychiatry has found it difficult to develop a nosology that allows for the targeted treatment of disorders of the mind. This article sets out a possible way forward: harnessing systems theory to provide the conceptual constraints needed to link clinical phenomena with neurobiology. This approach builds on the insight that the mind is a system which, to regulate its environment, needs to have a model of that environment and needs to update predictions about it using the rules of inductive logic. It can be shown that Bayesian inference can be reduced to updating beliefs based on precision-weighted prediction errors, where a prediction error is the difference between actual and predicted input, and precision is the confidence associated with the input prediction. Precision weighting of prediction errors entails that a given discrepancy between outcome and prediction means more, and leads to greater belief updates, the more confidently the prediction was made. This provides a conceptual framework linking clinical experience with the pathophysiology underlying disorders of the mind. Limitations of this approach are discussed and ways to work around them illustrated. Initial steps and possible future directions toward a nosology based on failures of precision weighting are discussed.

The basal ganglia include the striatum (caudate, putamen, and ventral striatum) which receive from all cortical areas, and which project via the globus pallidus and substantia nigra back to the neocortex. The basal ganglia are implicated in stimulus-response habit learning, which may be provided by a reinforcement learning signal received by dopamine neurons responding to reward prediction error. The dopamine neurons may receive reward-related information from the orbitofrontal cortex, via the ventral striatum and habenula. The network mechanisms in the basal ganglia implement selection of a single output for behaviour, which is highly adaptive, by mutual direct inhibition between neurons.

This chapter presents some useful ideas on how to encourage the process of learning. It illustrates that speeding up learning can occur if the learner becomes more sensitive to prediction errors. This chapter suggests that as the brain is presented with a prediction error, it tries to learn a generative model. It shows that the generative model’s trial-to-trial change is a forgetting rate. Presumably, biological learning is a process of state estimation, and a process in which the brain learns the structure of the generative model.

The idea that the brain is an organ for prediction error minimization is becoming increasingly influential. Since this idea posits action as playing a central role, it has the potential to reveal new perspectives on troubled notions of action, sense of agency, and body ownership. Elucidating these notions may help ascertain how close this new framework is to contemporary views of embodied, enactive, and extended cognition, which also places action central to cognition. The prediction error minimization framework suggests novel and somewhat provocative notions of action, sense of agency, and bodily ownership and, in important respects, it pulls in the opposite direction from the embodied, extended, and enactive approaches.

A brief concluding section summarizes the main points that have emerged from considering the predictive mind. The mind exists in prediction. Our perceptual experience of the world arises in our attempts at predicting our own current sensory input. This notion spreads to attention and agency. Perception, attention and agency are three different ways of doing the same thing: accounting for sensory input as well as we can from inside the confines of the skull. We are good at this, mostly, but it is a precarious and fragile process because we are hostages to our prior beliefs, our noisy brains, the uncertain sensory deliverances from the world, and to the brain’s urge to rid itself efficiently of prediction error. Intriguing approaches to a number of different mental phenomena arise from applying to them the simple idea that the brain does nothing but minimize its prediction error.

The chapter proposes that the three different statistics, social, vegetative, and natural, are matched and compared with each other; this results in what the theory of predictive coding describes as actual input (and ultimately as prediction error); that is, the neural activity change in the reward system during value assignment. Such prediction error as the matching and comparison between different stimuli and their respective statistics is possible, however, only on the basis of coding the differences between the different stimuli—that is, difference-based coding—rather than coding the stimuli themselves, or stimulus-based coding. Hence, the example of reward again demonstrates predictive coding to presuppose difference-based coding. Furthermore, the discussion of reward shows that the concept of the prediction error and more specifically the concept of the actual input needs to be specified (and elaborated) by the matching and comparison between different statistics: natural, social, and vegetative.

A regulatory battle for control ensues in the central nervous system following a mismatch between the current physiological state of an organism as mapped in viscerosensory brain regions and the predicted body state as computed in visceromotor control regions. The discrepancy between the predicted and current body state (i.e. the “somatic error”) signals a need for corrective action, motivating changes in both cognition and behavior. This chapter argues that anxiety disorders are fundamentally driven by somatic errors that fail to be adaptively regulated, leaving the organism in a state of dissonance where the predicted body state is perpetually out of line with the current body state. Repeated failures to quell somatic error can result in long-term changes to interoceptive circuitry within the brain. This chapter explores the neuropsychiatric sequelae that can emerge following chronic allostatic dysregulation of somatic errors and discusses novel therapies that might help to correct this dysregulation.

In the third chapter, reward dependent instrumental learning and its computational modelling is explained. Reward prediction errors are encoded by phasic dopamine release and specific paradigms including reversal learning are described with respect to clinical findings in different mental disorders. The chapter illustrates how computational approaches avoid relying exclusively on subjective reports of patients and instead correlate specific computational steps during reward related learning with their neurobiological correlates.

The idea that prediction error minimization is the central principle for the brain is introduced, the central argument in the book is described, and its three main parts are set out. Part 1 describes and explains the theory, drawing on work on computational neuroscience and cognitive science. Part 2 applies the theory to the way we manage to perceptually represent the world, and how we can get the world wrong. Part 3 discusses how the theory applies to aspects of the mind such as attention and consciousness, and looks at what it tells us about the nature of mind in the world. This introductory chapter also situates the theory in historical, philosophical and neuroscientific contexts.

The quest to understand the relationship between neural activity and behavior has been ongoing for well over a hundred years. Although research based on the stimulus-and-response approach to behavior, advocated by behaviorists, flourished during the last century, this view does not, by design, account for unobservable variables (e.g., mental states). Putting aside this approach, modern cognitive science, cognitive neuroscience, neuroeconomics, and behavioral economics have sought to explain this connection computationally. One major hurdle lies in the fact that we lack even a simple model of cognitive function. This chapter sketches an application that connects neuromodulator function to decision making and the valuation that underlies it. The nature of this hypothesized connection offers a fruitful platform to understand some of the informational aspects of dopamine function in the brain and how it exposes many different ways of understanding motivated choice.

Chapter 3 introduces the contours of REC’s positive program for relating to and allying with other major theories of cognition. In these efforts it aims to provide analyses and arguments designed to sanitize, strengthen, and unify existing representational and antirepresentational offerings. This theoretical work takes the form of RECtification—a process through which the target accounts of cognition are radicalized by analysis and argument, rendering them REC-friendly. This chapter shows how this process works in action by targeting Predictive Processing accounts of Cognition, or PPC. Some theorists have already argued that it is fruitful to combine PPC with E-theories of cognition. Yet they continue subscribe to a cognitivist reading of PPC. This chapter shows how an alliance between PPC with E-theories of cognition can only be properly forged, by giving the central ideas of PPC a REC rendering. It also shows why this crucial adjustment to PPC avoids crippling problems. This reveals why allying with REC is independently well motivated and theoretically beneficial.

Montague proposes a computational model to understand addiction. His key idea is a special kind of reward prediction error signal in addicts. In his comments, Yaffe discusses what Montague's work on the neuroscience of addiction does and does not show about moral responsibility. Sripada then outlines how additional deficits in reflective judgments of addicts might also be relevant to their moral responsibility. Montague replies by agreeing that we need a new generation of models to capture the kinds of considerations raised by his commentators.

The medial orbitofrontal cortex projects reward-related information to the pregenual cingulate cortex, and the lateral orbitofrontal cortex projects punishment and non-reward information to the supracallosal anterior cingulate cortex. These projections provide the reward outcome information needed for action-outcome goal value dependent instrumental learning by the cingulate cortex. The orbitofrontal cortex also projects reward-related information to the striatum for stimulus-response habit learning. Via the striatal route, and further in part via the habenula, the orbitofrontal cortex provides information about rewards and non-rewards that reached the brainstem dopamine neurons, some of which respond to positive reward prediction error, and the serotonin (5HT) neurons. The orbitofrontal cortex is therefore perhaps the key brain region involved in reward processing in the brain. The orbitofrontal cortex also has projections that can influence autonomic function, in part via the insula.